A 3d automatic segmentation method based on mathematical morphology for multiphoton images of melanocyte-keratinocyte coculture skin model
Résumé
Melanocyte-keratinocyte coculture models are interesting in vitro systems used to identify the de-pigmenting or pro-pigmenting potential of cosmetic ingredients. This potential can be estimated by calculating the melanin density inside this model. Multiphoton microscopy is a privileged microscopy technique for this kind of evaluation, thanks to its low invasiveness and appropriate contrast (time resolved two photon excited fluorescence) for melanin detection. On multiphoton images, the first necessary step to calculate melanin density is to delimitate the pixels where the tissue is located. In this paper, we proposed a tissue segmentation method based on mathematical morphology.
This pigmented coculture model contains two types of cells: keratinocytes and melanocytes that form a three dimensional tissue with a thickness of about 40 µm. The samples are reconstructed in 96 well plates, fixed in 4% formalin and rinsed in PBS prior to image acquisition. Multiphoton imaging was performed with a LEICA TCS SP8 microscopy at 760 nm, 40x/1.1NA W objective. A multiphoton 3D (x, y, z) image of 205x205x50 µm3 volume corresponds to a stack of 25 en face images of 512x512 pixels (0.4 µm/pixel) acquired with 2 µm z-step.
In this kind of models, segmentation is made difficult by the fact that the intensity of the fluorescence signal is heterogeneous over the tissue: dark regions inside the images can correspond either to background or to cytoplasmic regions. Therefore, the first step of our method consists in applying a horizontal area closing with size A. This operator fills all 2D dark structures which are smaller than A, closing any possible small and dark connections between the exterior and the interior of the tissue. Parameter A is taken equal to 150 µm2, i.e. the area of a cell. Afterwards, a reconstruction by erosion, starting from the first and last slides, is applied in 3D. This operation fills all dark structures inside the tissue. Finally, a simple threshold at the noise level produces the final mask of the tissue.
The method has been evaluated on a database containing 24 3D images, of which 4 had been manually segmented. The results were considered satisfactory by the experts.
This tissue segmentation method has been integrated in a software suite and is robust and fast enough in order to be used in an automatic process.
We are currently working on the following step, namely melanin quantification, to estimate the global melanin density, its z-distribution inside the tissue and localization in keratinocytes and melanocytes.